Research On Image Inpainting Based On Deep Learning

Image inpainting refers to the process of using undamaged information in the image to achieve image completion and get a better visual effect when the image information is lost.As an important image processing method,deep learning technology is also widely used in image inpainting.Existing image inpainting methods based on CNN and based on DCGAN cannot achieve the completion of damaged images with random damage locations and large damaged areas.In response to those situations,this thesis launched the research of deep learning technology in the field of image inpainting,designed and implemented an image inpaintting method based on Wasserstein and consistency,and realized the inpainting of damaged images with random damaged locations and large damaged areas.The main contents of this thesis are as follows:(1)Realize the image inpainting algorithm based on CNN and DCGAN,and research and analyze the experimental results..For images with random damage locations and large damaged areas,two inpainting network are designed and implemented.Through the analysis of respective image inpainting results,the advantages of CNN for image feature extraction and the defects of MSE loss function are clarified;Clarified the advantages of the DCGAN on quality of image generation and its performance problems in the training process.(2)Design and implement an image inpainting algorithm based on Wasserstein and consistency.In view of the performance of the DCGAN caused by the model crash during the training process,this thesis introduces WGAN to prevent this phenomenon;for the CNN in the image completion process due to the MSE loss function caused by the inpainting of the image blur problem,Combining the structural information prediction ability of the CNN with the adversarial strategy optimization ability of the discriminant network in the WGAN network,an image inpainting algorithm based on Wassserstein and consistency is constructed,which realizes the completion of damaged images in random damaged locations and larger damaged areas.(3)Test and verify the algorithm of this thesis launched.On the Celeb A data set,the inpainting quality of the algorithm launched in this thesis and other algorithms is tested and compared through subjective evaluation methods and objective evaluation methods.The comparison results show that the subjective evaluation scores and objective parameter indicators of the algorithm launched in this thesis are higher than those based on CNN and the repair methods based on DCGAN,indicating the effectiveness and feasibility of the algorithm in this thesis.In this thesis,an image inpainting algorithm based on Wasserstein and consistency is used to achieve the completion of random damaged locations and larger damaged areas.The inpainting image has a clear visual effect and has good look and feel conditions.The design and implementation of the algorithm in this thesis deepens the application of deep learning technology in the field of image inpainting,and to a certain extent promotes the development of image inpainting technology.